Avoid deadlock when starting table in attach thread of ReplicatedMergeTree

src/Storages/MergeTree/ReplicatedMergeTreeRestartingThread.h

@@ -29,6 +29,7 @@ class ReplicatedMergeTreeRestartingThread
 
     void shutdown(bool part_of_full_shutdown);
 
+    void run();
 private:
     StorageReplicatedMergeTree & storage;
     String log_name;
@@ -43,8 +44,6 @@ class ReplicatedMergeTreeRestartingThread
     UInt32 consecutive_check_failures = 0;  /// How many consecutive checks have failed
     bool first_time = true;                 /// Activate replica for the first time.
 
-    void run();
-
     /// Restarts table if needed, returns false if it failed to restart replica.
     bool runImpl();
 

src/Storages/StorageReplicatedMergeTree.cpp

@@ -4416,21 +4416,30 @@ void StorageReplicatedMergeTree::startupImpl(bool from_attach_thread)
 
         startBeingLeader();
 
-        /// In this thread replica will be activated.
-        restarting_thread.start();
+        if (from_attach_thread)
+        {
+            /// Try activating replica in current thread.
+            restarting_thread.run();
+        }
+        else
+        {
+            /// Activate replica in a separate thread.
+            restarting_thread.start();
+
+            /// Wait while restarting_thread finishing initialization.
+            /// NOTE It does not mean that replication is actually started after receiving this event.
+            /// It only means that an attempt to startup replication was made.
+            /// Table may be still in readonly mode if this attempt failed for any reason.
+            startup_event.wait();
+        }
+
         /// And this is just a callback
         session_expired_callback_handler = EventNotifier::instance().subscribe(Coordination::Error::ZSESSIONEXPIRED, [this]()
         {
             LOG_TEST(log, "Received event for expired session. Waking up restarting thread");
             restarting_thread.start();
         });
 
-        /// Wait while restarting_thread finishing initialization.
-        /// NOTE It does not mean that replication is actually started after receiving this event.
-        /// It only means that an attempt to startup replication was made.
-        /// Table may be still in readonly mode if this attempt failed for any reason.
-        startup_event.wait();
-
         startBackgroundMovesIfNeeded();
 
         part_moves_between_shards_orchestrator.start();

tests/integration/test_replicated_table_attach/configs/config.xml

@@ -0,0 +1,3 @@
+<clickhouse>
+    <background_schedule_pool_size>1</background_schedule_pool_size>
+</clickhouse>

tests/integration/test_replicated_table_attach/test.py

@@ -0,0 +1,60 @@
+import pytest
+
+from helpers.cluster import ClickHouseCluster
+from helpers.network import PartitionManager
+
+
+cluster = ClickHouseCluster(__file__)
+
+node = cluster.add_instance(
+    "node",
+    main_configs=["configs/config.xml"],
+    with_zookeeper=True,
+    stay_alive=True,
+)
+
+
+@pytest.fixture(scope="module")
+def started_cluster():
+    try:
+        cluster.start()
+        yield cluster
+
+    finally:
+        cluster.shutdown()
+
+
+def test_startup_with_small_bg_pool(started_cluster):
+    node.query(
+        "CREATE TABLE replicated_table (k UInt64, i32 Int32) ENGINE=ReplicatedMergeTree('/clickhouse/replicated_table', 'r1') ORDER BY k"
+    )
+
+    node.query("INSERT INTO replicated_table VALUES(20, 30)")
+
+    def assert_values():
+        assert node.query("SELECT * FROM replicated_table") == "20\t30\n"
+
+    assert_values()
+    node.restart_clickhouse(stop_start_wait_sec=10)
+    assert_values()
+
+    node.query("DROP TABLE replicated_table SYNC")
+
+
+def test_startup_with_small_bg_pool_partitioned(started_cluster):
+    node.query(
+        "CREATE TABLE replicated_table_partitioned (k UInt64, i32 Int32) ENGINE=ReplicatedMergeTree('/clickhouse/replicated_table_partitioned', 'r1') ORDER BY k"
+    )
+
+    node.query("INSERT INTO replicated_table_partitioned VALUES(20, 30)")
+
+    def assert_values():
+        assert node.query("SELECT * FROM replicated_table_partitioned") == "20\t30\n"
+
+    assert_values()
+    with PartitionManager() as pm:
+        pm.drop_instance_zk_connections(node)
+        node.restart_clickhouse(stop_start_wait_sec=20)
+        assert_values()
+
+    node.query("DROP TABLE replicated_table_partitioned SYNC")